We study the problem of distinguishing between two independent samples $\mathbf{G}_n^1,\mathbf{G}_n^2$ of a binomial random graph $G(n,p)$ by first order (FO) sentences. Shelah and Spencer proved that, for a constant $α\in(0,1)$, $G(n,n^{-α})$ obeys FO zero-one law if and only if $α$ is irrational. Therefore, for irrational $α\in(0,1)$, any fixed FO sentence does not distinguish between $\mathbf{G}_n^1,\mathbf{G}_n^2$ with asymptotical probability 1 (w.h.p.) as $n\to\infty$. We show that the minimum quantifier depth $\mathbf{k}_α$ of a FO sentence $\varphi=\varphi(\mathbf{G}_n^1,\mathbf{G}_n^2)$ distinguishing between $\mathbf{G}_n^1,\mathbf{G}_n^2$ depends on how closely $α$ can be approximated by rationals: (1) for all non-Liouville $α\in(0,1)$, $\mathbf{k}_α=Ω(\ln\ln\ln n)$ w.h.p.; (2) there are irrational $α\in(0,1)$ with $\mathbf{k}_α$ that grow arbitrarily slowly w.h.p.; (3) $\mathbf{k}_α=O_p(\frac{\ln n}{\ln\ln n})$ for all $α\in(0,1)$. The main ingredients in our proofs are a novel randomized algorithm that generates asymmetric strictly balanced graphs as well as a new method to study symmetry groups of randomly perturbed graphs.